whispercpp 1.3.3 → 1.3.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/ext/ruby_whisper_params.c +55 -25
- data/ext/sources/CMakeLists.txt +1 -1
- data/ext/sources/bindings/javascript/package.json +1 -1
- data/ext/sources/build-xcframework.sh +24 -0
- data/ext/sources/examples/CMakeLists.txt +1 -0
- data/ext/sources/examples/addon.node/addon.cpp +19 -19
- data/ext/sources/examples/addon.node/index.js +7 -5
- data/ext/sources/examples/bench/bench.cpp +26 -16
- data/ext/sources/examples/bench.wasm/index-tmpl.html +10 -9
- data/ext/sources/examples/cli/cli.cpp +4 -2
- data/ext/sources/examples/command/command.cpp +26 -24
- data/ext/sources/examples/command.wasm/index-tmpl.html +5 -4
- data/ext/sources/examples/common-ggml.cpp +2 -0
- data/ext/sources/examples/lsp/lsp.cpp +19 -17
- data/ext/sources/examples/server/server.cpp +24 -13
- data/ext/sources/examples/server.py +6 -1
- data/ext/sources/examples/stream/stream.cpp +4 -2
- data/ext/sources/examples/stream.wasm/emscripten.cpp +6 -6
- data/ext/sources/examples/stream.wasm/index-tmpl.html +82 -5
- data/ext/sources/examples/talk-llama/CMakeLists.txt +2 -2
- data/ext/sources/examples/talk-llama/llama-adapter.cpp +101 -4
- data/ext/sources/examples/talk-llama/llama-adapter.h +6 -0
- data/ext/sources/examples/talk-llama/llama-arch.cpp +588 -15
- data/ext/sources/examples/talk-llama/llama-arch.h +58 -1
- data/ext/sources/examples/talk-llama/llama-batch.cpp +103 -71
- data/ext/sources/examples/talk-llama/llama-batch.h +31 -18
- data/ext/sources/examples/talk-llama/llama-chat.cpp +120 -5
- data/ext/sources/examples/talk-llama/llama-chat.h +7 -0
- data/ext/sources/examples/talk-llama/llama-context.cpp +460 -357
- data/ext/sources/examples/talk-llama/llama-context.h +44 -29
- data/ext/sources/examples/talk-llama/llama-cparams.h +4 -4
- data/ext/sources/examples/talk-llama/llama-graph.cpp +543 -271
- data/ext/sources/examples/talk-llama/llama-graph.h +278 -168
- data/ext/sources/examples/talk-llama/llama-hparams.cpp +118 -4
- data/ext/sources/examples/talk-llama/llama-hparams.h +61 -15
- data/ext/sources/examples/talk-llama/llama-impl.h +2 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +326 -0
- data/ext/sources/examples/talk-llama/{llama-kv-cache-unified-iswa.h → llama-kv-cache-iswa.h} +38 -29
- data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +2020 -0
- data/ext/sources/examples/talk-llama/llama-kv-cache.h +358 -27
- data/ext/sources/examples/talk-llama/llama-kv-cells.h +80 -28
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +56 -36
- data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +30 -29
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +48 -19
- data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +13 -14
- data/ext/sources/examples/talk-llama/llama-memory.h +16 -10
- data/ext/sources/examples/talk-llama/llama-model-loader.cpp +2 -0
- data/ext/sources/examples/talk-llama/llama-model-loader.h +3 -2
- data/ext/sources/examples/talk-llama/llama-model.cpp +7165 -2336
- data/ext/sources/examples/talk-llama/llama-model.h +60 -9
- data/ext/sources/examples/talk-llama/llama-quant.cpp +48 -10
- data/ext/sources/examples/talk-llama/llama-sampling.cpp +226 -126
- data/ext/sources/examples/talk-llama/llama-vocab.cpp +440 -13
- data/ext/sources/examples/talk-llama/llama-vocab.h +45 -0
- data/ext/sources/examples/talk-llama/llama.cpp +65 -10
- data/ext/sources/examples/talk-llama/llama.h +95 -177
- data/ext/sources/examples/talk-llama/talk-llama.cpp +9 -6
- data/ext/sources/examples/talk-llama/unicode.cpp +207 -0
- data/ext/sources/examples/talk-llama/unicode.h +45 -0
- data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +4 -2
- data/ext/sources/examples/whisper.wasm/index-tmpl.html +17 -16
- data/ext/sources/ggml/CMakeLists.txt +59 -31
- data/ext/sources/ggml/cmake/ggml-config.cmake.in +132 -93
- data/ext/sources/ggml/include/ggml-backend.h +17 -1
- data/ext/sources/ggml/include/ggml-cpu.h +1 -1
- data/ext/sources/ggml/include/ggml-metal.h +1 -6
- data/ext/sources/ggml/include/ggml-opt.h +25 -6
- data/ext/sources/ggml/include/ggml-webgpu.h +19 -0
- data/ext/sources/ggml/include/ggml-zdnn.h +17 -0
- data/ext/sources/ggml/include/ggml.h +221 -16
- data/ext/sources/ggml/src/CMakeLists.txt +17 -2
- data/ext/sources/ggml/src/ggml-alloc.c +265 -141
- data/ext/sources/ggml/src/ggml-backend-impl.h +4 -1
- data/ext/sources/ggml/src/ggml-backend-reg.cpp +30 -13
- data/ext/sources/ggml/src/ggml-backend.cpp +221 -38
- data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
- data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +5 -4
- data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +14 -0
- data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +3 -1
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +903 -717
- data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +143 -25
- data/ext/sources/ggml/src/ggml-cann/common.h +143 -1
- data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +488 -69
- data/ext/sources/ggml/src/ggml-common.h +17 -0
- data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +40 -18
- data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +4 -2
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +132 -596
- data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +14 -286
- data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +103 -582
- data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +162 -589
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +265 -437
- data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +3 -58
- data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +521 -353
- data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +54 -314
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +184 -675
- data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +4679 -1657
- data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +32 -2
- data/ext/sources/ggml/src/ggml-cpu/common.h +14 -0
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -6
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +70 -42
- data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +35 -28
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +152 -18
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +7 -1
- data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +227 -97
- data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +474 -1116
- data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1587 -1177
- data/ext/sources/ggml/src/ggml-cpu/ops.h +5 -8
- data/ext/sources/ggml/src/ggml-cpu/quants.c +35 -0
- data/ext/sources/ggml/src/ggml-cpu/quants.h +8 -0
- data/ext/sources/ggml/src/ggml-cpu/repack.cpp +458 -47
- data/ext/sources/ggml/src/ggml-cpu/repack.h +22 -0
- data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +89 -60
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1024 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +13 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +3196 -0
- data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +26 -0
- data/ext/sources/ggml/src/ggml-cpu/traits.cpp +2 -2
- data/ext/sources/ggml/src/ggml-cpu/traits.h +1 -1
- data/ext/sources/ggml/src/ggml-cpu/vec.cpp +170 -26
- data/ext/sources/ggml/src/ggml-cpu/vec.h +506 -63
- data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +20 -16
- data/ext/sources/ggml/src/ggml-cuda/add-id.cu +58 -0
- data/ext/sources/ggml/src/ggml-cuda/add-id.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +330 -191
- data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/common.cuh +250 -63
- data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +1 -4
- data/ext/sources/ggml/src/ggml-cuda/conv2d.cu +166 -0
- data/ext/sources/ggml/src/ggml-cuda/conv2d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/convert.cu +95 -22
- data/ext/sources/ggml/src/ggml-cuda/convert.cuh +15 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
- data/ext/sources/ggml/src/ggml-cuda/cpy.cu +64 -307
- data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +2 -14
- data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +14 -40
- data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +498 -367
- data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +137 -91
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +755 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +3 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +593 -0
- data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +86 -50
- data/ext/sources/ggml/src/ggml-cuda/fattn.cu +185 -198
- data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
- data/ext/sources/ggml/src/ggml-cuda/getrows.cu +50 -39
- data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +379 -107
- data/ext/sources/ggml/src/ggml-cuda/im2col.cu +196 -35
- data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +1 -0
- data/ext/sources/ggml/src/ggml-cuda/mean.cu +56 -2
- data/ext/sources/ggml/src/ggml-cuda/mma.cuh +198 -45
- data/ext/sources/ggml/src/ggml-cuda/mmf.cu +123 -0
- data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +496 -0
- data/ext/sources/ggml/src/ggml-cuda/mmq.cu +206 -57
- data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +1262 -721
- data/ext/sources/ggml/src/ggml-cuda/{mmv.cu → mmvf.cu} +53 -53
- data/ext/sources/ggml/src/ggml-cuda/{mmv.cuh → mmvf.cuh} +3 -3
- data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +64 -73
- data/ext/sources/ggml/src/ggml-cuda/norm.cu +284 -12
- data/ext/sources/ggml/src/ggml-cuda/norm.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
- data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/pad.cu +46 -23
- data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
- data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/quantize.cu +12 -10
- data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +53 -0
- data/ext/sources/ggml/src/ggml-cuda/roll.cu +67 -0
- data/ext/sources/ggml/src/ggml-cuda/roll.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/rope.cu +21 -27
- data/ext/sources/ggml/src/ggml-cuda/scale.cu +14 -11
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +276 -0
- data/ext/sources/ggml/src/ggml-cuda/set-rows.cuh +7 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cu +34 -0
- data/ext/sources/ggml/src/ggml-cuda/softcap.cuh +5 -0
- data/ext/sources/ggml/src/ggml-cuda/softmax.cu +126 -59
- data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +10 -2
- data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +322 -100
- data/ext/sources/ggml/src/ggml-cuda/sum.cu +6 -10
- data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +21 -4
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +21 -18
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +259 -0
- data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +14 -0
- data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +3 -3
- data/ext/sources/ggml/src/ggml-cuda/unary.cu +90 -0
- data/ext/sources/ggml/src/ggml-cuda/unary.cuh +8 -0
- data/ext/sources/ggml/src/ggml-cuda/upscale.cu +92 -6
- data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +110 -22
- data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
- data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +58 -36
- data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -3
- data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +10 -2
- data/ext/sources/ggml/src/ggml-impl.h +119 -9
- data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -7
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +446 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.h +52 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +33 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +600 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +1376 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +226 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +1308 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +136 -63
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +3158 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +82 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +718 -0
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +2854 -1503
- data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +18 -8
- data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +18 -0
- data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +2510 -242
- data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +107 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +84 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/div.cl +66 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +370 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +370 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +373 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +27 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/glu.cl +177 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/group_norm.cl +49 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +1 -1
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +73 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +132 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +133 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +80 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +79 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +3 -2
- data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +189 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +34 -13
- data/ext/sources/ggml/src/ggml-opencl/kernels/sub.cl +66 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +20 -0
- data/ext/sources/ggml/src/ggml-opencl/kernels/tsembd.cl +2 -2
- data/ext/sources/ggml/src/ggml-opencl/kernels/upscale.cl +2 -3
- data/ext/sources/ggml/src/ggml-opt.cpp +97 -41
- data/ext/sources/ggml/src/ggml-quants.c +111 -16
- data/ext/sources/ggml/src/ggml-quants.h +6 -0
- data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +67 -47
- data/ext/sources/ggml/src/ggml-sycl/backend.hpp +2 -0
- data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +15 -5
- data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +6 -0
- data/ext/sources/ggml/src/ggml-sycl/concat.cpp +25 -16
- data/ext/sources/ggml/src/ggml-sycl/conv.cpp +10 -4
- data/ext/sources/ggml/src/ggml-sycl/convert.cpp +166 -99
- data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +72 -306
- data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +213 -1
- data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +67 -49
- data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +1 -31
- data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +79 -29
- data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
- data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +14 -26
- data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +9 -6
- data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +328 -323
- data/ext/sources/ggml/src/ggml-sycl/gla.cpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +2 -2
- data/ext/sources/ggml/src/ggml-sycl/mmq.cpp +80 -60
- data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +201 -132
- data/ext/sources/ggml/src/ggml-sycl/norm.cpp +74 -55
- data/ext/sources/ggml/src/ggml-sycl/quantize.hpp +133 -0
- data/ext/sources/ggml/src/ggml-sycl/quants.hpp +8 -9
- data/ext/sources/ggml/src/ggml-sycl/rope.cpp +35 -42
- data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +234 -0
- data/ext/sources/ggml/src/ggml-sycl/set_rows.hpp +8 -0
- data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +3 -3
- data/ext/sources/ggml/src/ggml-sycl/tsembd.cpp +12 -6
- data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +2 -6
- data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +16 -12
- data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3492 -883
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +41 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +13 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +39 -29
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +349 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +2 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +66 -12
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +154 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +2 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +6 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +4 -2
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +21 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +69 -24
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +60 -20
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +98 -42
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +64 -27
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +74 -13
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +39 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +4 -17
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +19 -10
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +25 -15
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +18 -14
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +126 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +65 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +11 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +140 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +144 -531
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +206 -38
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.comp +556 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +12 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +15 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +111 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +24 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +53 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +55 -11
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +1 -4
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.comp +5 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +1 -1
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +29 -7
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +4 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +17 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +38 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +4 -3
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +101 -9
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +69 -5
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp +25 -0
- data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +335 -77
- data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +54 -0
- data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1558 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add.tmpl.wgsl +44 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_in_place.tmpl.wgsl +41 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +45 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +930 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +60 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +124 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.tmpl.wgsl +874 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul.tmpl.wgsl +44 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_in_place.tmpl.wgsl +41 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +907 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +57 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_in_place.wgsl +48 -0
- data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +81 -0
- data/ext/sources/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
- data/ext/sources/ggml/src/ggml-zdnn/common.hpp +59 -0
- data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +628 -0
- data/ext/sources/ggml/src/ggml-zdnn/mmf.cpp +80 -0
- data/ext/sources/ggml/src/ggml-zdnn/mmf.hpp +12 -0
- data/ext/sources/ggml/src/ggml-zdnn/utils.cpp +79 -0
- data/ext/sources/ggml/src/ggml-zdnn/utils.hpp +19 -0
- data/ext/sources/ggml/src/ggml.c +478 -98
- data/ext/sources/ggml/src/gguf.cpp +8 -1
- data/ext/sources/src/whisper.cpp +23 -46
- data/ext/sources/tests/CMakeLists.txt +8 -1
- data/ext/sources/tests/test-vad-full.cpp +3 -3
- data/ext/sources/tests/test-vad.cpp +2 -2
- data/lib/whisper/model/uri.rb +1 -1
- data/sig/whisper.rbs +7 -0
- data/test/test_params.rb +8 -0
- data/test/test_whisper.rb +1 -1
- data/whispercpp.gemspec +1 -1
- metadata +164 -157
- data/ext/sources/examples/talk-llama/llama-kv-cache-unified-iswa.cpp +0 -279
- data/ext/sources/examples/talk-llama/llama-kv-cache-unified.cpp +0 -1841
- data/ext/sources/examples/talk-llama/llama-kv-cache-unified.h +0 -303
- data/ext/sources/ggml/include/ggml-kompute.h +0 -50
- data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +0 -107
- data/ext/sources/ggml/src/ggml-amx/common.h +0 -94
- data/ext/sources/ggml/src/ggml-amx/ggml-amx.cpp +0 -446
- data/ext/sources/ggml/src/ggml-amx/mmq.cpp +0 -2510
- data/ext/sources/ggml/src/ggml-amx/mmq.h +0 -17
- data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +0 -30
- data/ext/sources/ggml/src/ggml-cann/kernels/ascendc_kernels.h +0 -19
- data/ext/sources/ggml/src/ggml-cann/kernels/dup.cpp +0 -234
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f16.cpp +0 -197
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f32.cpp +0 -190
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -204
- data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -191
- data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -218
- data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -216
- data/ext/sources/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -295
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +0 -357
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +0 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +0 -365
- data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +0 -3
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +0 -482
- data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +0 -472
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
- data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
- data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +0 -166
- data/ext/sources/ggml/src/ggml-kompute/ggml-kompute.cpp +0 -2251
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +0 -112
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +0 -58
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +0 -25
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +0 -30
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +0 -22
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +0 -17
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +0 -31
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +0 -31
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +0 -38
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +0 -39
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +0 -44
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +0 -69
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +0 -51
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +0 -33
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +0 -35
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +0 -140
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +0 -106
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +0 -73
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +0 -28
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +0 -84
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +0 -21
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +0 -53
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +0 -52
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +0 -19
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +0 -23
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +0 -22
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +0 -72
- data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +0 -71
- data/ext/sources/ggml/src/ggml-metal/ggml-metal.m +0 -6280
|
@@ -0,0 +1,189 @@
|
|
|
1
|
+
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
|
|
2
|
+
|
|
3
|
+
kernel void kernel_set_rows_f32_i64(
|
|
4
|
+
global char * src0,
|
|
5
|
+
ulong offset0,
|
|
6
|
+
global char * src1,
|
|
7
|
+
ulong offset1,
|
|
8
|
+
global char * dst,
|
|
9
|
+
ulong offsetd,
|
|
10
|
+
int ne01,
|
|
11
|
+
ulong nb01,
|
|
12
|
+
ulong nb02,
|
|
13
|
+
ulong nb03,
|
|
14
|
+
int ne11,
|
|
15
|
+
int ne12,
|
|
16
|
+
ulong nb10,
|
|
17
|
+
ulong nb11,
|
|
18
|
+
ulong nb12,
|
|
19
|
+
int nblk0,
|
|
20
|
+
ulong nb1,
|
|
21
|
+
ulong nb2,
|
|
22
|
+
ulong nb3
|
|
23
|
+
) {
|
|
24
|
+
src0 = src0 + offset0;
|
|
25
|
+
src1 = src1 + offset1;
|
|
26
|
+
dst = dst + offsetd;
|
|
27
|
+
|
|
28
|
+
int i03 = get_group_id(2);
|
|
29
|
+
int i02 = get_group_id(1);
|
|
30
|
+
int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
|
|
31
|
+
|
|
32
|
+
if (i01 >= ne01) {
|
|
33
|
+
return;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
int i12 = i03%ne12;
|
|
37
|
+
int i11 = i02%ne11;
|
|
38
|
+
|
|
39
|
+
int i10 = i01;
|
|
40
|
+
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
|
|
41
|
+
|
|
42
|
+
global float * dst_row = (global float *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
|
|
43
|
+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
44
|
+
|
|
45
|
+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
|
|
46
|
+
dst_row[ind] = (float)src_row[ind];
|
|
47
|
+
}
|
|
48
|
+
}
|
|
49
|
+
|
|
50
|
+
kernel void kernel_set_rows_f16_i64(
|
|
51
|
+
global char * src0,
|
|
52
|
+
ulong offset0,
|
|
53
|
+
global char * src1,
|
|
54
|
+
ulong offset1,
|
|
55
|
+
global char * dst,
|
|
56
|
+
ulong offsetd,
|
|
57
|
+
int ne01,
|
|
58
|
+
ulong nb01,
|
|
59
|
+
ulong nb02,
|
|
60
|
+
ulong nb03,
|
|
61
|
+
int ne11,
|
|
62
|
+
int ne12,
|
|
63
|
+
ulong nb10,
|
|
64
|
+
ulong nb11,
|
|
65
|
+
ulong nb12,
|
|
66
|
+
int nblk0,
|
|
67
|
+
ulong nb1,
|
|
68
|
+
ulong nb2,
|
|
69
|
+
ulong nb3
|
|
70
|
+
) {
|
|
71
|
+
src0 = src0 + offset0;
|
|
72
|
+
src1 = src1 + offset1;
|
|
73
|
+
dst = dst + offsetd;
|
|
74
|
+
|
|
75
|
+
int i03 = get_group_id(2);
|
|
76
|
+
int i02 = get_group_id(1);
|
|
77
|
+
int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
|
|
78
|
+
|
|
79
|
+
if (i01 >= ne01) {
|
|
80
|
+
return;
|
|
81
|
+
}
|
|
82
|
+
|
|
83
|
+
int i12 = i03%ne12;
|
|
84
|
+
int i11 = i02%ne11;
|
|
85
|
+
|
|
86
|
+
int i10 = i01;
|
|
87
|
+
long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
|
|
88
|
+
|
|
89
|
+
global half * dst_row = (global half *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
|
|
90
|
+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
91
|
+
|
|
92
|
+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
|
|
93
|
+
dst_row[ind] = src_row[ind];
|
|
94
|
+
}
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
kernel void kernel_set_rows_f32_i32(
|
|
98
|
+
global char * src0,
|
|
99
|
+
ulong offset0,
|
|
100
|
+
global char * src1,
|
|
101
|
+
ulong offset1,
|
|
102
|
+
global char * dst,
|
|
103
|
+
ulong offsetd,
|
|
104
|
+
int ne01,
|
|
105
|
+
ulong nb01,
|
|
106
|
+
ulong nb02,
|
|
107
|
+
ulong nb03,
|
|
108
|
+
int ne11,
|
|
109
|
+
int ne12,
|
|
110
|
+
ulong nb10,
|
|
111
|
+
ulong nb11,
|
|
112
|
+
ulong nb12,
|
|
113
|
+
int nblk0,
|
|
114
|
+
ulong nb1,
|
|
115
|
+
ulong nb2,
|
|
116
|
+
ulong nb3
|
|
117
|
+
) {
|
|
118
|
+
src0 = src0 + offset0;
|
|
119
|
+
src1 = src1 + offset1;
|
|
120
|
+
dst = dst + offsetd;
|
|
121
|
+
|
|
122
|
+
int i03 = get_group_id(2);
|
|
123
|
+
int i02 = get_group_id(1);
|
|
124
|
+
int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
|
|
125
|
+
|
|
126
|
+
if (i01 >= ne01) {
|
|
127
|
+
return;
|
|
128
|
+
}
|
|
129
|
+
|
|
130
|
+
int i12 = i03%ne12;
|
|
131
|
+
int i11 = i02%ne11;
|
|
132
|
+
|
|
133
|
+
int i10 = i01;
|
|
134
|
+
int i1 = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
|
|
135
|
+
|
|
136
|
+
global float * dst_row = (global float *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
|
|
137
|
+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
138
|
+
|
|
139
|
+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
|
|
140
|
+
dst_row[ind] = (float)src_row[ind];
|
|
141
|
+
}
|
|
142
|
+
}
|
|
143
|
+
|
|
144
|
+
kernel void kernel_set_rows_f16_i32(
|
|
145
|
+
global char * src0,
|
|
146
|
+
ulong offset0,
|
|
147
|
+
global char * src1,
|
|
148
|
+
ulong offset1,
|
|
149
|
+
global char * dst,
|
|
150
|
+
ulong offsetd,
|
|
151
|
+
int ne01,
|
|
152
|
+
ulong nb01,
|
|
153
|
+
ulong nb02,
|
|
154
|
+
ulong nb03,
|
|
155
|
+
int ne11,
|
|
156
|
+
int ne12,
|
|
157
|
+
ulong nb10,
|
|
158
|
+
ulong nb11,
|
|
159
|
+
ulong nb12,
|
|
160
|
+
int nblk0,
|
|
161
|
+
ulong nb1,
|
|
162
|
+
ulong nb2,
|
|
163
|
+
ulong nb3
|
|
164
|
+
) {
|
|
165
|
+
src0 = src0 + offset0;
|
|
166
|
+
src1 = src1 + offset1;
|
|
167
|
+
dst = dst + offsetd;
|
|
168
|
+
|
|
169
|
+
int i03 = get_group_id(2);
|
|
170
|
+
int i02 = get_group_id(1);
|
|
171
|
+
int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
|
|
172
|
+
|
|
173
|
+
if (i01 >= ne01) {
|
|
174
|
+
return;
|
|
175
|
+
}
|
|
176
|
+
|
|
177
|
+
int i12 = i03%ne12;
|
|
178
|
+
int i11 = i02%ne11;
|
|
179
|
+
|
|
180
|
+
int i10 = i01;
|
|
181
|
+
int i1 = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
|
|
182
|
+
|
|
183
|
+
global half * dst_row = (global half *) (dst + i1*nb1 + i02*nb2 + i03*nb3);
|
|
184
|
+
global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
185
|
+
|
|
186
|
+
for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
|
|
187
|
+
dst_row[ind] = src_row[ind];
|
|
188
|
+
}
|
|
189
|
+
}
|
|
@@ -22,32 +22,49 @@
|
|
|
22
22
|
REQD_SUBGROUP_SIZE_64
|
|
23
23
|
#endif
|
|
24
24
|
kernel void kernel_soft_max_4_f16(
|
|
25
|
-
global
|
|
25
|
+
global char * src0,
|
|
26
26
|
ulong offset0,
|
|
27
|
-
global
|
|
27
|
+
global char * src1,
|
|
28
28
|
ulong offset1,
|
|
29
|
-
global
|
|
29
|
+
global char * src2,
|
|
30
|
+
ulong offset2,
|
|
31
|
+
global char * dst,
|
|
30
32
|
ulong offsetd,
|
|
31
33
|
int ne00,
|
|
32
|
-
|
|
33
|
-
|
|
34
|
+
ulong nb01,
|
|
35
|
+
ulong nb02,
|
|
36
|
+
ulong nb03,
|
|
37
|
+
int ne12,
|
|
38
|
+
int ne13,
|
|
39
|
+
ulong nb11,
|
|
40
|
+
ulong nb12,
|
|
41
|
+
ulong nb13,
|
|
42
|
+
ulong nb1,
|
|
43
|
+
ulong nb2,
|
|
44
|
+
ulong nb3,
|
|
34
45
|
float scale,
|
|
35
46
|
float max_bias,
|
|
36
47
|
float m0,
|
|
37
48
|
float m1,
|
|
38
49
|
int n_head_log2
|
|
39
50
|
) {
|
|
40
|
-
src0 =
|
|
41
|
-
src1 =
|
|
42
|
-
|
|
51
|
+
src0 = src0 + offset0;
|
|
52
|
+
src1 = src1 + offset1;
|
|
53
|
+
src2 = src2 + offset2;
|
|
54
|
+
dst = dst + offsetd;
|
|
43
55
|
|
|
44
56
|
int i03 = get_group_id(2);
|
|
45
57
|
int i02 = get_group_id(1);
|
|
46
58
|
int i01 = get_group_id(0);
|
|
47
59
|
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
60
|
+
int i13 = i03%ne13;
|
|
61
|
+
int i12 = i02%ne12;
|
|
62
|
+
int i11 = i01;
|
|
63
|
+
|
|
64
|
+
global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
65
|
+
global half4 * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
|
|
66
|
+
global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
|
|
67
|
+
global float4 * pdst4 = (global float4 *)(dst + i01*nb1 + i02*nb2 + i03*nb3);
|
|
51
68
|
|
|
52
69
|
float slope = 1.0f;
|
|
53
70
|
|
|
@@ -62,7 +79,7 @@ kernel void kernel_soft_max_4_f16(
|
|
|
62
79
|
}
|
|
63
80
|
|
|
64
81
|
// parallel max
|
|
65
|
-
float4 lmax4 = -INFINITY;
|
|
82
|
+
float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
|
|
66
83
|
for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
|
|
67
84
|
lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
|
|
68
85
|
}
|
|
@@ -79,7 +96,11 @@ kernel void kernel_soft_max_4_f16(
|
|
|
79
96
|
}
|
|
80
97
|
float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
|
|
81
98
|
|
|
82
|
-
|
|
99
|
+
float sum = sub_group_reduce_add(lsum);
|
|
100
|
+
|
|
101
|
+
if (psrc2) {
|
|
102
|
+
sum += exp(psrc2[i02] - max);
|
|
103
|
+
}
|
|
83
104
|
|
|
84
105
|
for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
|
|
85
106
|
pdst4[i00] /= sum;
|
|
@@ -22,32 +22,49 @@
|
|
|
22
22
|
REQD_SUBGROUP_SIZE_64
|
|
23
23
|
#endif
|
|
24
24
|
kernel void kernel_soft_max_4(
|
|
25
|
-
global
|
|
25
|
+
global char * src0,
|
|
26
26
|
ulong offset0,
|
|
27
|
-
global
|
|
27
|
+
global char * src1,
|
|
28
28
|
ulong offset1,
|
|
29
|
-
global
|
|
29
|
+
global char * src2,
|
|
30
|
+
ulong offset2,
|
|
31
|
+
global char * dst,
|
|
30
32
|
ulong offsetd,
|
|
31
33
|
int ne00,
|
|
32
|
-
|
|
33
|
-
|
|
34
|
+
ulong nb01,
|
|
35
|
+
ulong nb02,
|
|
36
|
+
ulong nb03,
|
|
37
|
+
int ne12,
|
|
38
|
+
int ne13,
|
|
39
|
+
ulong nb11,
|
|
40
|
+
ulong nb12,
|
|
41
|
+
ulong nb13,
|
|
42
|
+
ulong nb1,
|
|
43
|
+
ulong nb2,
|
|
44
|
+
ulong nb3,
|
|
34
45
|
float scale,
|
|
35
46
|
float max_bias,
|
|
36
47
|
float m0,
|
|
37
48
|
float m1,
|
|
38
49
|
int n_head_log2
|
|
39
50
|
) {
|
|
40
|
-
src0 =
|
|
41
|
-
src1 =
|
|
42
|
-
|
|
51
|
+
src0 = src0 + offset0;
|
|
52
|
+
src1 = src1 + offset1;
|
|
53
|
+
src2 = src2 + offset2;
|
|
54
|
+
dst = dst + offsetd;
|
|
43
55
|
|
|
44
56
|
int i03 = get_group_id(2);
|
|
45
57
|
int i02 = get_group_id(1);
|
|
46
58
|
int i01 = get_group_id(0);
|
|
47
59
|
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
60
|
+
int i13 = i03%ne13;
|
|
61
|
+
int i12 = i02%ne12;
|
|
62
|
+
int i11 = i01;
|
|
63
|
+
|
|
64
|
+
global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
65
|
+
global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
|
|
66
|
+
global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
|
|
67
|
+
global float4 * pdst4 = (global float4 *)(dst + i01*nb1 + i02*nb2 + i03*nb3);
|
|
51
68
|
|
|
52
69
|
float slope = 1.0f;
|
|
53
70
|
|
|
@@ -62,7 +79,7 @@ kernel void kernel_soft_max_4(
|
|
|
62
79
|
}
|
|
63
80
|
|
|
64
81
|
// parallel max
|
|
65
|
-
float4 lmax4 = -INFINITY;
|
|
82
|
+
float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
|
|
66
83
|
for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
|
|
67
84
|
lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
|
|
68
85
|
}
|
|
@@ -79,7 +96,11 @@ kernel void kernel_soft_max_4(
|
|
|
79
96
|
}
|
|
80
97
|
float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
|
|
81
98
|
|
|
82
|
-
|
|
99
|
+
float sum = sub_group_reduce_add(lsum);
|
|
100
|
+
|
|
101
|
+
if (psrc2) {
|
|
102
|
+
sum += exp(psrc2[i02] - max);
|
|
103
|
+
}
|
|
83
104
|
|
|
84
105
|
for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
|
|
85
106
|
pdst4[i00] /= sum;
|
|
@@ -22,32 +22,49 @@
|
|
|
22
22
|
REQD_SUBGROUP_SIZE_64
|
|
23
23
|
#endif
|
|
24
24
|
kernel void kernel_soft_max_f16(
|
|
25
|
-
global
|
|
25
|
+
global char * src0,
|
|
26
26
|
ulong offset0,
|
|
27
|
-
global
|
|
27
|
+
global char * src1,
|
|
28
28
|
ulong offset1,
|
|
29
|
-
global
|
|
29
|
+
global char * src2,
|
|
30
|
+
ulong offset2,
|
|
31
|
+
global char * dst,
|
|
30
32
|
ulong offsetd,
|
|
31
33
|
int ne00,
|
|
32
|
-
|
|
33
|
-
|
|
34
|
+
ulong nb01,
|
|
35
|
+
ulong nb02,
|
|
36
|
+
ulong nb03,
|
|
37
|
+
int ne12,
|
|
38
|
+
int ne13,
|
|
39
|
+
ulong nb11,
|
|
40
|
+
ulong nb12,
|
|
41
|
+
ulong nb13,
|
|
42
|
+
ulong nb1,
|
|
43
|
+
ulong nb2,
|
|
44
|
+
ulong nb3,
|
|
34
45
|
float scale,
|
|
35
46
|
float max_bias,
|
|
36
47
|
float m0,
|
|
37
48
|
float m1,
|
|
38
49
|
int n_head_log2
|
|
39
50
|
) {
|
|
40
|
-
src0 =
|
|
41
|
-
src1 =
|
|
42
|
-
|
|
51
|
+
src0 = src0 + offset0;
|
|
52
|
+
src1 = src1 + offset1;
|
|
53
|
+
src2 = src2 + offset2;
|
|
54
|
+
dst = dst + offsetd;
|
|
43
55
|
|
|
44
56
|
int i03 = get_group_id(2);
|
|
45
57
|
int i02 = get_group_id(1);
|
|
46
58
|
int i01 = get_group_id(0);
|
|
47
59
|
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
60
|
+
int i13 = i03%ne13;
|
|
61
|
+
int i12 = i02%ne12;
|
|
62
|
+
int i11 = i01;
|
|
63
|
+
|
|
64
|
+
global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
65
|
+
global half * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
|
|
66
|
+
global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
|
|
67
|
+
global float * pdst = (global float *)(dst + i01*nb1 + i02*nb2 + i03*nb3);
|
|
51
68
|
|
|
52
69
|
float slope = 1.0f;
|
|
53
70
|
|
|
@@ -62,7 +79,7 @@ kernel void kernel_soft_max_f16(
|
|
|
62
79
|
}
|
|
63
80
|
|
|
64
81
|
// parallel max
|
|
65
|
-
float lmax = -INFINITY;
|
|
82
|
+
float lmax = psrc2 ? psrc2[i02] : -INFINITY;
|
|
66
83
|
for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
|
|
67
84
|
lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
|
|
68
85
|
}
|
|
@@ -78,7 +95,11 @@ kernel void kernel_soft_max_f16(
|
|
|
78
95
|
pdst[i00] = exp_psrc0;
|
|
79
96
|
}
|
|
80
97
|
|
|
81
|
-
|
|
98
|
+
float sum = sub_group_reduce_add(lsum);
|
|
99
|
+
|
|
100
|
+
if (psrc2) {
|
|
101
|
+
sum += exp(psrc2[i02] - max);
|
|
102
|
+
}
|
|
82
103
|
|
|
83
104
|
for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
|
|
84
105
|
pdst[i00] /= sum;
|
|
@@ -22,32 +22,49 @@
|
|
|
22
22
|
REQD_SUBGROUP_SIZE_64
|
|
23
23
|
#endif
|
|
24
24
|
kernel void kernel_soft_max(
|
|
25
|
-
global
|
|
25
|
+
global char * src0,
|
|
26
26
|
ulong offset0,
|
|
27
|
-
global
|
|
27
|
+
global char * src1,
|
|
28
28
|
ulong offset1,
|
|
29
|
-
global
|
|
29
|
+
global char * src2,
|
|
30
|
+
ulong offset2,
|
|
31
|
+
global char * dst,
|
|
30
32
|
ulong offsetd,
|
|
31
33
|
int ne00,
|
|
32
|
-
|
|
33
|
-
|
|
34
|
+
ulong nb01,
|
|
35
|
+
ulong nb02,
|
|
36
|
+
ulong nb03,
|
|
37
|
+
int ne12,
|
|
38
|
+
int ne13,
|
|
39
|
+
ulong nb11,
|
|
40
|
+
ulong nb12,
|
|
41
|
+
ulong nb13,
|
|
42
|
+
ulong nb1,
|
|
43
|
+
ulong nb2,
|
|
44
|
+
ulong nb3,
|
|
34
45
|
float scale,
|
|
35
46
|
float max_bias,
|
|
36
47
|
float m0,
|
|
37
48
|
float m1,
|
|
38
49
|
int n_head_log2
|
|
39
50
|
) {
|
|
40
|
-
src0 =
|
|
41
|
-
src1 =
|
|
42
|
-
|
|
51
|
+
src0 = src0 + offset0;
|
|
52
|
+
src1 = src1 + offset1;
|
|
53
|
+
src2 = src2 + offset2;
|
|
54
|
+
dst = dst + offsetd;
|
|
43
55
|
|
|
44
56
|
int i03 = get_group_id(2);
|
|
45
57
|
int i02 = get_group_id(1);
|
|
46
58
|
int i01 = get_group_id(0);
|
|
47
59
|
|
|
48
|
-
|
|
49
|
-
|
|
50
|
-
|
|
60
|
+
int i13 = i03%ne13;
|
|
61
|
+
int i12 = i02%ne12;
|
|
62
|
+
int i11 = i01;
|
|
63
|
+
|
|
64
|
+
global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
|
|
65
|
+
global float * pmask = src1 != src0 ? (global float *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
|
|
66
|
+
global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
|
|
67
|
+
global float * pdst = (global float *)(dst + i01*nb1 + i02*nb2 + i03*nb3);
|
|
51
68
|
|
|
52
69
|
float slope = 1.0f;
|
|
53
70
|
|
|
@@ -62,7 +79,7 @@ kernel void kernel_soft_max(
|
|
|
62
79
|
}
|
|
63
80
|
|
|
64
81
|
// parallel max
|
|
65
|
-
float lmax = -INFINITY;
|
|
82
|
+
float lmax = psrc2 ? psrc2[i02] : -INFINITY;
|
|
66
83
|
for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
|
|
67
84
|
lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
|
|
68
85
|
}
|
|
@@ -78,7 +95,11 @@ kernel void kernel_soft_max(
|
|
|
78
95
|
pdst[i00] = exp_psrc0;
|
|
79
96
|
}
|
|
80
97
|
|
|
81
|
-
|
|
98
|
+
float sum = sub_group_reduce_add(lsum);
|
|
99
|
+
|
|
100
|
+
if (psrc2) {
|
|
101
|
+
sum += exp(psrc2[i02] - max);
|
|
102
|
+
}
|
|
82
103
|
|
|
83
104
|
for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
|
|
84
105
|
pdst[i00] /= sum;
|
|
@@ -70,3 +70,69 @@ kernel void kernel_sub_row(
|
|
|
70
70
|
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
|
|
71
71
|
dst[gid] = src0[gid] - src1[idx1];
|
|
72
72
|
}
|
|
73
|
+
|
|
74
|
+
kernel void kernel_sub_f16(
|
|
75
|
+
global char * src0,
|
|
76
|
+
ulong offset0,
|
|
77
|
+
global char * src1,
|
|
78
|
+
ulong offset1,
|
|
79
|
+
global char * dst,
|
|
80
|
+
ulong offsetd,
|
|
81
|
+
ulong nb00,
|
|
82
|
+
ulong nb01,
|
|
83
|
+
ulong nb02,
|
|
84
|
+
ulong nb03,
|
|
85
|
+
int ne10,
|
|
86
|
+
int ne11,
|
|
87
|
+
int ne12,
|
|
88
|
+
int ne13,
|
|
89
|
+
ulong nb10,
|
|
90
|
+
ulong nb11,
|
|
91
|
+
ulong nb12,
|
|
92
|
+
ulong nb13,
|
|
93
|
+
int ne0,
|
|
94
|
+
ulong nb0,
|
|
95
|
+
ulong nb1,
|
|
96
|
+
ulong nb2,
|
|
97
|
+
ulong nb3
|
|
98
|
+
) {
|
|
99
|
+
src0 = src0 + offset0;
|
|
100
|
+
src1 = src1 + offset1;
|
|
101
|
+
dst = dst + offsetd;
|
|
102
|
+
|
|
103
|
+
int i03 = get_group_id(2);
|
|
104
|
+
int i02 = get_group_id(1);
|
|
105
|
+
int i01 = get_group_id(0);
|
|
106
|
+
|
|
107
|
+
int i13 = i03 % ne13;
|
|
108
|
+
int i12 = i02 % ne12;
|
|
109
|
+
int i11 = i01 % ne11;
|
|
110
|
+
|
|
111
|
+
global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
|
|
112
|
+
global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
|
|
113
|
+
global char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1;
|
|
114
|
+
|
|
115
|
+
for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
|
|
116
|
+
const int i10 = i0 % ne10;
|
|
117
|
+
*((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) - *((global half *)(src1_ptr + i10*nb10));
|
|
118
|
+
}
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
kernel void kernel_sub_row_f16(
|
|
122
|
+
global half4 * src0,
|
|
123
|
+
ulong offset0,
|
|
124
|
+
global half4 * src1,
|
|
125
|
+
ulong offset1,
|
|
126
|
+
global half4 * dst,
|
|
127
|
+
ulong offsetd,
|
|
128
|
+
int ne
|
|
129
|
+
) {
|
|
130
|
+
src0 = (global half4*)((global char*)src0 + offset0);
|
|
131
|
+
src1 = (global half4*)((global char*)src1 + offset1);
|
|
132
|
+
dst = (global half4*)((global char*)dst + offsetd);
|
|
133
|
+
|
|
134
|
+
// This performs better than using %.
|
|
135
|
+
uint gid = get_global_id(0);
|
|
136
|
+
uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
|
|
137
|
+
dst[gid] = src0[gid] - src1[idx1];
|
|
138
|
+
}
|
|
@@ -24,6 +24,26 @@ kernel void kernel_transpose_16(
|
|
|
24
24
|
write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
|
|
25
25
|
}
|
|
26
26
|
|
|
27
|
+
// Padded kernel for irregular shape
|
|
28
|
+
kernel void kernel_transpose_16_4x1(
|
|
29
|
+
__read_only image1d_buffer_t input,
|
|
30
|
+
__write_only image1d_buffer_t output,
|
|
31
|
+
const uint rows,
|
|
32
|
+
const uint cols
|
|
33
|
+
) {
|
|
34
|
+
|
|
35
|
+
const int i = get_global_id(0);
|
|
36
|
+
const int j = get_global_id(1);
|
|
37
|
+
const int j_2 = j << 2;
|
|
38
|
+
|
|
39
|
+
half temp0 = read_imageh(input, (j_2 + 0) * cols + i).x;
|
|
40
|
+
half temp1 = read_imageh(input, (j_2 + 1) * cols + i).x;
|
|
41
|
+
half temp2 = read_imageh(input, (j_2 + 2) * cols + i).x;
|
|
42
|
+
half temp3 = read_imageh(input, (j_2 + 3) * cols + i).x;
|
|
43
|
+
|
|
44
|
+
write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
|
|
45
|
+
}
|
|
46
|
+
|
|
27
47
|
// 32-bit transpose, loading/storing a 4x4 tile of elements
|
|
28
48
|
kernel void kernel_transpose_32(
|
|
29
49
|
__read_only image1d_buffer_t input,
|
|
@@ -26,8 +26,8 @@ kernel void kernel_timestep_embedding(
|
|
|
26
26
|
local_half_dim = logical_dim / 2;
|
|
27
27
|
local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
|
|
28
28
|
|
|
29
|
-
if (logical_dim % 2 != 0 && local_j ==
|
|
30
|
-
local_embed_data_ptr[
|
|
29
|
+
if (logical_dim % 2 != 0 && local_j == local_half_dim) {
|
|
30
|
+
local_embed_data_ptr[2 * local_half_dim] = 0.0f;
|
|
31
31
|
}
|
|
32
32
|
|
|
33
33
|
if (local_j >= local_half_dim) {
|
|
@@ -60,7 +60,8 @@ kernel void kernel_upscale_bilinear(
|
|
|
60
60
|
float sf0,
|
|
61
61
|
float sf1,
|
|
62
62
|
float sf2,
|
|
63
|
-
float sf3
|
|
63
|
+
float sf3,
|
|
64
|
+
float pixel_offset
|
|
64
65
|
) {
|
|
65
66
|
global const char * src_base = (global const char *)p_src0 + off_src0;
|
|
66
67
|
global float * dst_base = (global float *)((global char *)p_dst + off_dst);
|
|
@@ -80,8 +81,6 @@ kernel void kernel_upscale_bilinear(
|
|
|
80
81
|
int i02_src = (int)(i12_dst / sf2);
|
|
81
82
|
int i03_src = (int)(i13_dst / sf3);
|
|
82
83
|
|
|
83
|
-
const float pixel_offset = 0.5f;
|
|
84
|
-
|
|
85
84
|
float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
|
|
86
85
|
long y0_src = (long)floor(y_src_f);
|
|
87
86
|
long y1_src = y0_src + 1;
|